Research Papers & Studies¶
Key research papers, studies, and reports referenced across Blueprint for an AI-First Company. Organized by topic area with source attribution where known.
Human-AI Collaboration¶
Studies on how humans and AI work together effectively -- and when collaboration fails. Central to Chapter 2 (The AI-First Mindset) and the Human-AI Collaboration framework.
| Study | Organization | Key Finding | Topic |
|---|---|---|---|
| Systematic review of human-AI teaming outcomes | MIT Center for Collective Intelligence | Reviewed 100+ studies and found that human-AI combinations don't, on average, outperform the best human-only or AI-only systems. Collaboration succeeds under three specific conditions. | Human-AI collaboration effectiveness |
| Human-AI teaming study | Nature Human Behaviour | Identified conditions where human-AI teams succeed: tasks where humans outperform AI alone, content creation tasks, and creation tasks specifically involving generative AI. | Conditions for effective collaboration |
| COHUMAIN framework | Carnegie Mellon University | Proposed viewing AI as a "partner under direction" rather than a teammate -- capable of strengthening capabilities but not replacing human judgment. | AI collaboration design |
| AI coding productivity study | METR (Model Evaluation and Threat Research) | Randomized controlled trial found developers were 19% slower with AI assistance but predicted they'd be 24% faster -- a 43-point perception gap. | Developer productivity measurement |
AI Adoption & Business Impact¶
Large-scale surveys and analyses on AI adoption patterns across industries. Referenced throughout the book, particularly in chapters on strategy, operations, and teams.
| Study | Organization | Key Finding | Topic |
|---|---|---|---|
| Global Survey on AI | McKinsey & Company | 88% of companies fail with AI implementation. | AI implementation failure rates |
| AI adoption studies | MIT Sloan | Identified the Buy, Boost, or Build framework for AI capability development. BCG deployed AI to consultants and reduced interview processing from two weeks to 2-3 days. | AI adoption strategies |
| AI productivity research | Accenture | Trained all 700,000 employees in agentic AI. Marketing department saw 25% external brand value improvement and nearly one-third reduction in manual tasks. | Enterprise AI training at scale |
| Enterprise AI spending survey | Multiple sources (industry surveys) | Average monthly AI spend jumped from $62,964 in 2024 to $85,521 in 2025. Companies spending over $100K monthly increased from 20% to 45%. | AI cost trends |
| AI pilot failure analysis | MIT (NANDA study) | 95% of enterprise AI pilots fail to reach production with measurable value. The root cause is infrastructure and process failures, not model quality. | Pilot-to-production gap |
| AI skills and productivity | AWS | AI skills boost productivity by at least 39% across organizations surveyed. | Workforce productivity |
| Enterprise AI tool survey | Zapier | Only 35% of enterprise AI tools go through proper approval channels. 28% of enterprises use more than 10 different AI apps. 76% experienced negative outcomes from disconnected AI. | AI tool sprawl |
AI Risk & Safety¶
Research on AI risks, security vulnerabilities, and safety frameworks. Core to Chapter 11 (Ethics, Governance, and Risk) and the 7 AI Risks and Mitigations framework.
| Study | Organization | Key Finding | Topic |
|---|---|---|---|
| AI Risk Management Framework (AI RMF) | NIST (National Institute of Standards and Technology) | Four-function framework -- Govern, Map, Measure, Manage -- for identifying, assessing, and mitigating AI risks systematically. | AI risk management |
| DeepSeek security evaluation | Cisco | Found a 100% attack success rate against DeepSeek-R1 -- the model failed to block any harmful prompts. Generates insecure code at four times the rate of competitors. | AI model security |
| AI-generated code security | Multiple research groups | Only 55% of AI-generated code is secure. XSS vulnerabilities appear 86% of the time. SQL injection in 20% of cases. | Code security |
| AI incident tracking | Industry tracking | AI incidents jumped 21% from 2024 to 2025 as companies expanded autonomy without expanding controls. | AI incident trends |
| Model degradation study | ML research | 91% of ML models experience performance degradation without intervention due to data drift and concept drift. | Model maintenance |
| AI hiring discrimination study | Research (2024) | AI screening favored white applicants over Black applicants with identical credentials 85% of the time. | Algorithmic bias |
| Secrets leakage report | Industry report (2024) | 23.77 million secrets were leaked through AI systems -- a 25% increase from the prior year. | Data security |
Data Strategy & Competitive Advantage¶
Research on data flywheels, data moats, and AI economics. Referenced in Chapter 9 (Data Strategy) and the Data Flywheel and Data Moats frameworks.
| Study | Organization | Key Finding | Topic |
|---|---|---|---|
| Model collapse study | Nature | Models trained on AI-generated content exhibit "narrower range of output over time" -- each generation drifts further from reality. | Synthetic data risks |
| AI startup failure analysis | Industry research | AI startups face a 92% failure rate within 18 months. 42% of companies abandoned most AI initiatives in 2025. | AI startup survival |
| Data accuracy trends | Industry survey | Data accuracy in the U.S. declined from 63.5% in 2021 to just 26.6% in 2024. | Data quality |
| AI wrapper economics | Industry analysis | Average AI wrapper operates at 25-60% gross margins versus 70-90% for traditional SaaS. 60-70% of AI wrappers generate zero revenue. | AI business economics |
| AI customer service effectiveness | Qualtrics (2025) | AI-powered customer service fails at four times the rate of other AI tasks. Nearly one in five consumers see no benefits from AI customer service. | Customer experience |
AI Governance & Regulation¶
Regulatory frameworks and governance research. Referenced in Chapter 11 and the AI Governance Framework.
| Framework / Study | Organization | Description | Topic |
|---|---|---|---|
| EU AI Act | European Union | Tiered penalty regime: fines up to EUR 35M or 7% of global turnover for prohibited practices, 3% for high-risk violations. | AI regulation |
| ISO 42001 | ISO | International standard for AI management systems, providing a framework for responsible AI development and deployment. | AI standards |
| AI board oversight study | Financial services industry | 84% increase in board oversight disclosure around AI in 2024 across financial services. | Corporate governance |
| AI Red Team taxonomy | Microsoft | Formalized taxonomy for AI agent failures published April 2025, aligning with the seven failure modes discussed in the book. | AI safety testing |
| AI Fairness 360 | IBM | Open source toolkit for detecting and mitigating bias in AI models across the machine learning lifecycle. | Bias detection |
| Fairlearn | Microsoft | Open source toolkit for assessing and improving fairness of AI systems. | Fairness in AI |
Related Frameworks¶
- Human-AI Collaboration -- Applies the MIT and Nature Human Behaviour findings
- 7 AI Risks and Mitigations -- Operationalizes the NIST AI RMF research
- Data Flywheel -- Built on the data strategy research
- AI Governance Framework -- Implements EU AI Act and ISO 42001 requirements
- 8 Patterns for AI Coding -- Informed by the METR developer productivity study